high performance computing on graphics processing units: hgpu.org

With the rise of heterogeneous chips available in the market, where integrated GPU cores and CPU cores reside in the same chip and share a unified memory, it is possible to have better execution schemes for many graph algorithms. Graph algorithms can exhibit producer-consumer behavior, a varying amount of parallelism during execution, and irregularity which results in inefficiency. The inefficiency problem could be solved by exploiting heterogeneity between cores. In this work, I provide an understanding of the executions of some graph algorithms in heterogeneous chips and accelerate their executions by using fine-grain software optimization techniques. To achieve this, I introduce two different fine-grain execution techniques to accelerate the Maximal Independent Set and Preflow-push graph algorithms, and present an evaluation of the techniques on a heterogeneous chip. My techniques, namely Overlapping Threads with Hot-Vertices and Task Switcher, provide 1.3x to 16x speedup over CPU-only execution depending on the input and the algorithm.